Research On Steel Surface Defect Detect Method Based On One-stage Detector

Steel surface defect detection based on convolutional neural network has become a research hotspot in the industry at present.At present,the object detection framework used in industry is the one-stage detector framework.The one-stage detector has fewer parameters and is fast enough to meet the demand for speed in industrial production,but these object detectors are generally trained based on a data set with balanced datasets.Therefore,in this paper,we will use the best-performing one-stage detector,YOLOv3,as the base object detection framework,and improve YOLOv3 for the actual steel surface defect data to improve the classes imbalance and scale imbalance problems existing in the dataset.Firstly,the data augmentation algorithm of perceptual hash clustering(PHC)and online hard sample moderate mining algorithm(OHEMM)are proposed for the classes imbalance problem.the PHC method will cluster the images in the original dataset according to the features,so that the augmented data features are comprehensive.The OHEMM method,based on the OHEM,proposes the idea of moderate mining of hard samples,limiting the number of hard samples mined to prevent overfitting of the model.The method is integrated on the YOLOv3 and conducted on the Northeastern University steel surface defect dataset(NET-DET)for classes imbalance experiments,and the experimental results show that the method improves the accuracy of the YOLOv3 object detection framework.Secondly,the Mask algorithm is proposed for the scale imbalance problem.Scale imbalance can lead to difficulties in bounding box prediction.The method adds a prediction branch to the model,and the predicted generated Mask can be corrected for the predicted bounding box to improve the accuracy of bounding box prediction.The labels used for training Mask prediction are generated based on the bounding box labels in the dataset,and no additional labeling is required.The method is integrated on the YOLOv3 and scale imbalance experiments are conducted on the NET-DET,and the results show that the method improves the scale imbalance problem.Then,in this paper,the methods used to solve the above problems are fused in two to improve the simultaneous classes imbalance and scale imbalance in steel surface defect data.In this paper,these algorithms are integrated simultaneously on the YOLOv3 object detection framework according to the characteristics of the different algorithms to verify the stack ability of the algorithms.Tested on the NET-DET dataset,the experimental results show that the methods used to solve each problem can be fused with each other and integrated on the object detection framework,and the experimental results are better than those of a single method.Finally,we summarize our work and present a few research directions in the future.